# 2026-03-04

- Added task `task-20260304-211216-acp-claude-codex` to `memory/tasks.json`:
  - Fix ACP runtime failures for Claude Code and Codex agents (`acpx` exit codes 1/5).
- Implemented immediate runtime mitigation for LiteLLM route reliability:
  - Updated default model from `litellm/copilot-gpt-5.3-codex` to `litellm/gpt-4o` in `/home/openclaw/.openclaw/openclaw.json`.

- LiteLLM/OpenClaw runtime root cause identified and fixed:
  - `~/.openclaw/agents/main/agent/models.json` had stale `litellm.baseUrl` set to `http://localhost:4000`.
  - Updated `litellm.baseUrl` to `http://192.168.153.113:18804/v1` to match active provider endpoint.
  - Post-fix validation runs for `litellm/gpt-4o` succeeded.
- User requested default model rollback after fix:
  - Default model set to `openai-codex/gpt-5.3-codex` in `~/.openclaw/openclaw.json`.
- Operational note:
  - Multiple delayed subagent completion events produced repeated `ok` messages; should batch/ignore stale completion spam in future updates.

- BOOT.md startup check executed (~2026-03-04T21:38Z) and state files updated:
  - Wrote `memory/startup-health.json` and `memory/boot-last-run.json`.
  - Gateway exposure confirmed `local-only` (loopback bind).
  - Backup signal fresh (`Backup complete` ~3.35h old at check time).
  - Update status: `up-to-date`.
- Startup security findings to follow up:
  - `openclaw security audit --deep` reported **1 critical / 2 warn** related to plugin posture:
    - critical: `acpx.bak` plugin flagged for dangerous code pattern,
    - warn: `plugins.allow` not explicitly set,
    - warn: extension tools reachable under permissive policy.
  - Recorded in `memory/startup-health.md`; overall startup status set to `warn`.
- Added TODO `task-20260304-2215-subagent-reliability` (high priority):
  - Fix subagent reliability and completion-event spam handling.
  - User requested explicit ownership and follow-through on subagent/tooling stability.
- Implemented "real fix" local TUI patch to suppress internal runtime context noise:
  - Edited `/home/openclaw/.npm-global/lib/node_modules/openclaw/dist/tui-LeOEBhMz.js`.
  - Added detection for internal runtime completion blocks (`OpenClaw runtime context (internal)` + `[Internal task completion event]`) and return empty text for those records in `extractTextFromMessage`.
  - Goal: hide internal orchestration handoff messages from TUI chat log while keeping normal user/assistant messages visible.
- Updated TODO `task-20260304-2215-subagent-reliability` to `in-progress` with patch note.
- Implemented upstream-source fix in cloned OpenClaw repo:
  - Repo: `external/openclaw-upstream`
  - Branch: `fix/tui-hide-internal-runtime-context`
  - Commit: `0f66a4547`
  - Change: in `src/tui/tui-formatters.ts`, suppress rendering of internal runtime completion context blocks in `extractTextFromMessage` when both markers are present.
  - Tests: updated `src/tui/tui-formatters.test.ts` with positive/negative coverage and ran `pnpm test:fast` successfully (suite passed).
- Added durable workflow guidance for OpenClaw code contributions and update hygiene:
  - Updated `AGENTS.md` with required upstream workflow + branch policy + update-safe notes.
  - Added `scripts/openclaw-upstream-workflow.md` with step-by-step commands (sync main, branch, test, PR, and post-update sync).
- Added explicit Git policy updates for OpenClaw upstream workflow:
  - Rebase-first sync (`git pull --rebase`) for upstream clone maintenance.
  - Rebase onto latest `origin/main` before push/PR.
  - Preferred PR merge strategy: **Rebase and merge**.
- Added explicit post-merge branch cleanup policy to workflow docs:
  - After rebase-merge PRs, delete feature/fix branches both remote and local.
  - Added command examples in `scripts/openclaw-upstream-workflow.md`.
- Added `scripts/openclaw-upstream-preflight.sh` and wired docs to enforce upstream fetch/sync before coding.
  - Script syncs `external/openclaw-upstream` (`fetch --all --prune`, `checkout main`, `pull --rebase`) and can create/rebase target branch.